Almost nine out of every ten dollars that American enterprises spend on open-source large models now goes to a Chinese vendor. That is the central finding of Menlo Ventures’ 2026 State of Enterprise AI report, based on a survey of close to 400 US enterprise AI decision-makers.
The split
Broken down, Moonshot’s Kimi takes 61 per cent of that spending. DeepSeek takes 15 per cent, Zhipu’s GLM 11 per cent and Alibaba’s Qwen 3 per cent. The best-placed Western model, Meta’s Llama, holds 7 per cent.
Menlo’s own reading is blunt. As Chinese models defined the open-source frontier, they took 89 per cent of open-source spending.
Price fell, volume exploded
The average price of an open-model token fell by two thirds over the year, from USD 0.45 per million to USD 0.13. Volume over the same period rose 22.5-fold, reaching 2.9 quadrillion tokens a month.
The result is a strange split in the market. Open models account for 19 per cent of platform token volume but only 4 per cent of enterprise model spending. They are used heavily and paid for lightly, which is exactly what a commodity layer looks like.
Usage rankings tell a second story
Menlo measures where the money goes. OpenRouter measures where the requests go, and the ordering is different. In the week from 26 October to 1 November, DeepSeek held 29.8 per cent of text requests across all model authors, up 7 percentage points. OpenAI was second at 18.8 per cent. Google continued to slide from 14.4 per cent. Z.ai, the international brand of Zhipu, was fourth at 11.9 per cent and Qwen sixth at 5.6 per cent.
That week was the 27th consecutive week in which Chinese models served more tokens than American ones. Among individual models, four of the top five by weekly volume were Chinese.
DeepSeek’s own numbers are large. Its V4.1 Flash model processed 114 trillion tokens on OpenRouter in the past 30 days, up 95 per cent month on month. The 1017 build pushed to 6 trillion tokens a day. Its desktop agent product Harness, live since 29 September, reports 1.54 million daily active users and 92 trillion tokens a day.
The budget line behind it
Menlo puts US enterprise AI deployment spending at USD 176.8 billion in 2026, up from USD 33 billion a year earlier, a 5.4-fold increase. Fifty-five per cent of the new budget is net new money rather than a shift out of software or headcount.
So the spend is real and it is accelerating. The part that should worry a procurement team is that the cheapest, most-used layer of that stack is now supplied overwhelmingly by vendors in a jurisdiction that US policy treats as a competitor. Nobody decided this in a meeting. It accumulated one API call at a time.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1098112884_120321309.